agentmark
Pass
Audited by Gen Agent Trust Hub on May 20, 2026
Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill makes frequent use of the
agentmarkCLI vianpxto perform project operations such as running experiments, generating types, and managing deployments. These operations utilize the vendor's specialized tooling.\n- [EXTERNAL_DOWNLOADS]: The skill is configured to fetch documentation and metadata fromdocs.agentmark.co(e.g.,llms.txt,mcpserver reference). These are trusted official sources for the AgentMark ecosystem.\n- [DATA_EXFILTRATION]: The skill provides instructions for linking local projects to AgentMark Cloud (app.agentmark.co), which involves the legitimate transmission of local traces and telemetry for observability and management purposes as part of the tool's core functionality.\n- [PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it is designed to ingest and process content from.prompt.mdxand.jsonlfiles, which may contain untrusted data.\n - Ingestion points: Dataset rows in
.jsonlfiles and prompt templates in.prompt.mdxfiles (specificallyworkflows/building-datasets.mdandworkflows/creating-prompts.md).\n - Boundary markers: Employs TemplateDX tags (e.g.,
<System>,<User>,<Assistant>) to provide structure and separation between instruction blocks, though it relies on correctly defined schemas to mitigate malicious property injection.\n - Capability inventory: The agent is empowered to execute CLI commands (
npx agentmark), conduct network operations to cloud endpoints, and modify local project files.\n - Sanitization: Relies on frontmatter-defined JSON Schemas for input validation as described in
workflows/creating-prompts.md.
Audit Metadata